Speaker
Description
Inspecting subsea oil and gas pipelines is expensive. Autonomous Underwater Vehicles have limited mission time, yet most operators still follow fixed inspection schedules that treat every pipeline section the same regardless of age, water depth, or operating history. Low-risk sections get checked anyway. High-risk sections wait their turn. This paper builds a digital twin framework that uses historical offshore incident data and machine learning to direct Autonomous Underwater Vehicle inspection toward the sections most likely to fail. The subsea network is modelled as a graph: nodes are junctions, manifolds, and riser bases; edges are pipeline sections carrying length, diameter, water depth, age, and operating pressure as attributes. Offshore incident records from the United States Pipeline and Hazardous Materials Safety Administration database train three failure-prediction models Logistic Regression as a baseline, Random Forest, and Extreme Gradient Boosting. Whichever model performs best feeds its risk predictions back onto the network graph, forming the digital twin. Shapley Additive Explanations are applied throughout, so every risk score can be traced back to the features driving it rather than treated as a black box. An Autonomous Underwater Vehicle agent is simulated on this graph under three inspection policies: random, scheduled round-robin, and machine learning–driven risk-prioritized routing. Detection rate, mean time to detection, and coverage efficiency are compared across all three. The entire framework runs on open data and free software, with no specialised hardware required built for offshore operators who cannot afford commercial digital twin platforms.
Keywords: digital twin, machine learning, predictive maintenance, subsea pipelines
| Topic Area | Smart Operations and Maintenance |
|---|